Files
pipelines/app/notes/read-aloud.md
T
alan 6367cce68a Read chat replies aloud with word highlighting
An assistant reply gets Read aloud beside Copy; while it reads it becomes
Pause/Resume and Stop, one reply at a time app-wide. The spoken word is
highlighted in the rendered reply (across multi-part replies) and the
transcript scrolls only when that word leaves the viewport. Reading stops
when its chat closes or is swapped, another reply starts, or the app goes
to the background. Markdown is spoken as sentences; code blocks are
announced as "Code block.".

Speech goes through a SpeechEngine seam with a flutter_tts implementation
(Android, macOS, web; hidden on Linux) and a fake for tests. Settings has
a Read aloud section: a voice picker grouped by language (device language
first, a sample per voice), a rate slider and reset, saved per device.

flutter_tts has no Swift package, so macOS/iOS gain the CocoaPods
integration (Podfiles with the Runner deployment targets); Android
declares the TTS_SERVICE query. The capture fixture gains ?speech=fake.
2026-11-13 01:29:04 +00:00

2.8 KiB

Read aloud

An assistant reply has Read aloud beside Copy (same IconButton style). While that reply reads it becomes Pause/Resume and Stop; one reply reads at a time app-wide. The spoken word is highlighted (theme primaryContainer) in the rendered reply and the transcript scrolls only when that word leaves the viewport. Settings > Read aloud picks the voice (grouped by language, device language first, a sample per voice), the rate (0.5x to 2.0x) and resets both; they are saved on the device.

Structure

  • lib/features/chat/speech/speech_engine.dart: the SpeechEngine seam (speak, pause, stop, setVoice, setRate, voices, start/pause/resume/ progress/completion/cancel/error callbacks).
  • flutter_tts_speech_engine.dart: the flutter_tts implementation. The rate is a multiplier (1.0 = normal); Apple platforms get half of it on their 0..1 scale. Linux reports unsupported, so the button and section are hidden.
  • speech_controller.dart: the app's single reader (speechController, a fake in tests via debugSpeechController). State changes notify the controller; the spoken word is the separate word notifier. Stops on hidden, paused or detached (not inactive, which is only a lost focus).
  • speech_document.dart: Markdown to sentences, with each spoken character's part and source offset. Code blocks (``` and ~~~) are "Code block.", tables "Table." then cells; rules are skipped.
  • chat_markdown_speech.dart: highlight spans and the reveal scroll for ChatMarkdownView.
  • Reader ids are <sessionId>/<row id> so a chat that closes or is swapped for another stops only its own reply (stopScope).

Platform notes

  • Resume speaks the paused text again; Android, Apple and web plugins all treat that as a resume. Android then reports progress against the remaining suffix, which the controller maps back to the whole text.
  • After a stop the controller waits (at most 400 ms) for the platform cancel, so a late cancel cannot end the next reply and web speech is idle before the next utterance.
  • Apple and web flutter_tts cannot clear a chosen voice. After a voice was set in a run, "System default" uses the device language's first voice.
  • Platforms without word boundaries read without a highlight.
  • flutter_tts has no Swift package, so macOS (and iOS) builds use CocoaPods for it: Podfile and the Pods-Runner xcconfig includes are the generated plugin integration.

Verification

Tests (fake engine): test/features/chat/speech/, test/features/chat/ui/transcript/chat_read_aloud_button_test.dart, chat_markdown_read_aloud_test.dart, test/features/chat/conversation/chat_thread_read_aloud_test.dart, test/features/settings/settings_read_aloud_section_test.dart.

Not verified on a device: real speech output, word boundaries and voices on Android, macOS and browsers.